Dynamic Selection of Fitness Function (DSF) for feature selection in software defect prediction

In this study, we consider three fitness functions, Accuracy, F Measure,and Voting and Validation, and use them to select features in twenty-four software datasets for feature selection. Based on the performance of a classifier on these selected features, using the DSF algorithm we select the fitness function which is most likely to correctly predict the defect susceptibility of a new class intance.

Dataset used

All datasets for the purpose of this project have been taken from the PROMISE repository of open source software- http://openscience.us/repo/defect

Using this repository

dataset/dataset contains all the preprocessed datasets which are input to DSE pipeline
dataset/annotated contains output of first fold of cassification, including results of all three base ensembles and best ensemle for it
dataset/DSF contains additional columns containing predictions of DSE and probabilities
dataset/metrics includes performance metrics of all base ensmebles on softwares
results contain performance metrics of DSF